Technology companies are actively seeking new ways to leverage artificial intelligence (AI) technologies, leading to the emergence of AI agents, which are not only able to generate content but are also able to take action based on the information they gain from their environment.
According to the Financial Times (FT), OpenAI CFO Sarah Friar predicts that AI agents will become the buzzword in 2025. The first batch will be deployed as researchers or assistants to help people manage daily tasks.
"I think 2025 is going to be the year that agentic systems finally hit the mainstream," OpenAI's new chief product officer Kevin Weil, said at a press event ahead of the company's annual Dev Day, according to The Verge.
What are AI agents?
AI agents take generative AI (GenAI) to a new level, can independently perform tasks, make decisions, and learn from their experiences.
According to Adnan Ijaz, director of product management for Amazon Q Developer, and Yoon Kim, an assistant professor at MIT's Computer Science and Artificial Intelligence Laboratory. AI agents usually follow a three-part workflow:
1. The user sets a goal and provides a prompt.
2. AI agents approach the task by breaking it down into smaller, simpler subtasks and collecting the needed data.
3. AI agents execute tasks using what's contained in their knowledge base plus the data they've amassed, making use of any functions they can call or tools they have at their disposal.
For instance, if the user requests to book a flight, the AI agent will first search for all flights that meet the specified criteria, then select the most affordable option, and proceed to make the booking through the airline's application programming interface (API).
However, humans can still guide the process and intervene when required. For example, if the cheapest flight has no available seats, the AI agent will notify the user, allowing them to decide on the next step.
AI Agents are poised to become the next golden goose
Jared Friedman, a partner at the Silicon Valley startup accelerator Y Combinator, believes that AI agents have the potential to surpass SaaS, with significant advantages in reducing costs while boosting efficiency.
Friedman advises entrepreneurs to identify highly repetitive and tedious administrative tasks and leverage AI agents to automate these processes.
AI agent tool use
In fact, several tech companies have already shifted their business focus and conducting preliminary tests of related functionalities in 2024.
Google made a surprise move just before the year-end holidays by releasing the first model of its Gemini 2.0 series, Gemini 2.0 Flash, and plans to expand the availability of Gemini 2.0 into Google's products by 2025.
In addition, Google showcased several AI agent prototypes built on Gemini 2.0, including Project Astra, Project Mariner, and Jules. Project Astra has the ability to converse in multiple languages and execute multi-step online tasks; Jules is designed to assist developers with coding tasks.
Microsoft
Microsoft announced the launch of a series of pre-built AI agents within Microsoft 365 at Ignite 2024. For instance, the project manager agent in Planner automatically assigns tasks, tracks progress, and sends notifications, while the coordinator agent in Teams provides real-time meeting notes and shares key information summaries.
Users can also create custom agents using the low-code platform, Copilot Studio.
Salesforce
SaaS company Salesforce launched its AI agent platform, Agentforce, in October. During its most recent earnings call, Salesforce revealed that it had secured multiple orders and plans to hire over 1,000 new employees to drive sales of the platform.
Startups are also pouring resources into R&D to capture market share.
OpenAI
In October, OpenAI introduced Swarm, an experimental multi-agent coordination framework, with its design centered around two core concepts: "routine" and "handoff." The former refers to a group of agents following instructions to complete specific tasks, while the latter allows for seamless transitions between agents specialized in distinct functions.
For example, in a customer service system, a triage agent conducts an initial assessment and forwards specific queries to agents with expertise in sales, support services, or refunds.
Furthermore, OpenAI is preparing to launch an AI agent called 'Operator' in January 2025, granting researchers access to its API, Bloomberg reported.
Perplexity
In early December 2024, Perplexity launched the "Buy With Pro" AI shopping tool, designed to assist users in searching for products online, comparing prices, and completing purchases directly.
Perplexity has also partnered with digital payment provider Stripe, which offers the AI agent a one-time charge card for online payments. This approach allows the AI agent to complete transactions without needing access to the user's bank account, thereby reducing potential risks.
However, the service is still facing challenges, including slow processing speeds and occasional transaction failures, requiring human intervention to ensure accuracy.
In October 2024, Anthropic introduced a "computer use" feature, based on the upgraded Claude 3.5 Sonnet, enabling AI to interpret screen displays and control computers like a human.
Challenges to adoption
David Singleton, former VP of Engineering at Google Android, established a company called /dev/agents, focused on creating operating systems for AI agents.
Singleton pointed out that the current difficulty in developing AI agents is too high, and he aims to build an AI agent development platform similar to Android, simplifying the development process for developers.
Beyond technical challenges, data privacy and security represent major obstacles for AI agents. These agents will access, analyze, and collect vast amounts of personal data. If targeted by malicious actors, personal information could be exposed and compromised.
Moreover, since AI agents are based on large language models (LLMs), hallucinations are a concern. Additionally, with AI agents handling tasks directly for consumers and bypassing the need to visit websites, this could disrupt the revenue streams and business models of retailers, publishers, and advertisers.
As AI agents interact more frequently with one another and humans are increasingly excluded, ensuring trust and accountability will be top of the concerns.